改进U-Net模型的隧道掌子面图像语义分割研究OA北大核心

The structural characteristics of the rock mass exposed at the tunnel face provide a direct ba-sis for assessing geotechnical conditions,formulating construction and support strategies,and mitigat-ing risks of accidents such as collapses and water inrush.When applying the U-Net model to the seg-mentation and recognition of tunnel face rock mass structure images,the downsampling process can lead to the loss of fine details in the rock mass,while the skip connections used during upsampling to transfer low-level features to higher levels may cause excessively large feature maps.To address these issues,an improved U-Net model is proposed by incorporating the Atrous Spatial Pyramid Pooling(ASPP)module and the Convolutional Block Attention Module(CBAM).Specifically,the ASPP is integrated into the skip connections of the U-Net model to capture multi-scale contextual information through atrous convolutions with varying dilation rates,enabling the fusion of features from diverse re-ceptive fields for a more comprehensive understanding of image content.Concurrently,the CBAM is embedded into the downsampling process of the U-Net model to enhancing the network focus more on useful features,thereby enhancing the representation capability of the extracted features.Experimental results demonstrate that the improved network model significantly outperforms the original U-Net in both segmentation and recognition performance.Evaluated on a tunnel face rock mass image dataset from a specific engineering project,the improved model achieves a Precision of 93.04%,mean Inter-section over Union(mIoU)of 74.98%,and a mean Pixel Accuracy(mPA)of 78.89%.